Data-driven deep learning model of shield vertical attitude prediction

It is difficult to control the vertical attitude during shield tunneling, and the shield often deviates from the design axis. In order to solve the problem that the existing shield attitude prediction models cannot accurately extract data features and effectively remove data noise, the time series information of shield tunneling measured data was fully exploited, and according to the shield tunneling project in Gengyun Road Station—Qingtan Road Station section of Hefei Metro Line 7, the collected tunneling parameters including removing the shutdown status data and abnormal data were preprocessed, a CNN-LSTM combined model for predicting the shield vertical attitude was proposed. The prediction performance of the model on the test set was compared with that of the traditional regression model. Finally, the performance of the model with different sample sizes and fixed network parameters was studied. The results show that the CNN-LSTM combined model achieves good performance in predicting the shield vertical attitude, which exhibits low mean absolute error EMA and root mean square error ERMS and a high coefficient of determination R2 on the test set, indicating small prediction errors and high prediction accuracy. Compared to the ARIMA, LSTM, and SVR models, the CNN-LSTM model improves the predicted R2 values by 1.04%, 19.75% and 79.63% on the test set, respectively. Furthermore, the model shows low EMA and ERMS, and the training time is significantly reduced. The performance of the CNN-LSTM model is influenced by different sample sizes. When the ratio of the training set sample size to the test set sample size is 4:1, the model achieves the highest predicted R2, indicating optimal prediction performance. Increasing the training set sample size may lead to overfitting and decrease the generalization ability of the model, while reducing the training set sample size may result in underfitting and decrease the prediction accuracy on the test set. Fixing certain network parameters can effectively reduce the number of training parameters and training time as well as improve the prediction accuracy of the model. The model performs the best when fixing the parameters of a 4-layer network, with a predicted R2 of 0.93, and predicted EMA and ERMS of 0.029 and 0.048, respectively. The training time is 46 s. © 2024 Central South University of Technology. All rights reserved.